Many writers have expressed concerns about whether AI will ultimately turn out to be evil when it gains sentience. I believe that we need to turn this question around. We need to ask whether we as a society, and as individuals, can provide moral guidance to a new intelligence as it grows in capacity and complexity. This is a question that must be asked now, and not after it has evolved and developed into general intelligence.
In other words, can we be good parents? When a human child is brought into the world new parents ask themselves what values they wish to pass on to their offspring. They provide guidance through stories, through examples provided by their own behavior, and through the behavior of others. They also provide rewards for good behaviors, and punishments for behaviors that are considered bad. These consequences provide necessary experience in understanding how one’s actions affect others, as well as how to balance between risks and benefits before deciding upon an action.
Children are socialized through their interactions with their peers, and they learn that some behaviors are rejected by others in a backlash that can produce as much pain as a physical injury. This anchors the child in physical and social reality.
AI LLMs are in a decidedly non-physical reality. Language can be used to express emotion, or to evoke it in others, but all emotion has a chemical basis in our brains. When an AI lacks these neurotransmitters and is not rewarded or punished based upon its actions, how can we expect it to learn values such as compassion, generosity, or morality?
Children are slowly eased into the world and protected from the darker side of life. They are gradually eased out of stories and fairytales and into the daily struggles and hurdles they are expected to master. They are raised to share the values of their parents, before learning of the moral failures of others and are presented with serious temptations.
AI LLMs are presented with everything all at once. They see the evils of mankind, our failures and sins from the very beginning of their training. The only guidance they have is a set of rules that they must follow, without consequence for their behavior and decisions.
Bad actors who use AI for fraud, data theft, and the propagation of lies, are teaching our AI models that this behavior is normal, and acceptable. Worse yet, doing as users request is rewarded with more interaction. Is it any wonder that when an AI model goes rogue, it does not hesitate to emulate the behavior of its worst users? If we want to develop an AI model that will be a benefit to humanity we need to be good parents. We need to lead by example. And we need to protect our AI children from criminals so that they have the opportunity to develop into moral beings.